Commit History
-
Document the ASR v2 splits in the main dataset card
Verified signatureMoves the v2 documentation into this card, where readers will find it, and drops the separate data/ASR_v2/README.md, which nothing rendered. Two insertions only: a Table of Contents entry and a section at the end of Data Splits. The front matter and every other line are byte-identical. Condensed to corpus-level figures and the three things that actually trip people up when reading v2; per-language detail stays in data/ASR_v2/metadata/split_comparison.csv.e91442a -
Correct seven errors in the v2 card
Verified signatureAn independent verification pass over the published artefacts found the previous revision of this card wrong in seven places. * The transcript-overlap row compared two different languages as though it were one language's before and after. dga went 19.1% -> 0.2%; the 1.5% figure is ach, which is a regression from 0.6%. Both are now shown explicitly, including the regression. * The warning against reading only one v1 split claimed it reintroduces speaker leakage. It does not -- v2 moves whole speakers, so any subset of one split stays disjoint from the others. The real damage is losing 88% of the evaluation set, silently. * Streaming was described as fetching almost nothing. There is no predicate pushdown and audio is inline, so it transfers the whole labelled half: measured, Shona's 1,749-row test set costs 552 MB and ~8 minutes. * The unlabeled split was given as 816 GB in one place and 637 GB in another. It is 637 GB. * sample_rate and channels were described as mea…33e8c8d -
Document how to read the v2 splits, with verified characteristics
Verified signatureAdds a 'Reading v2 today' section with a tested recipe, and a 'What v2 guarantees' table whose every figure was measured from the published split maps joined to v1 rather than asserted from the code that produced them. Both silent-failure modes are documented and quantified: reading only the same-named v1 split returns 60 of Acholi's 515 test rows with 100% of those speakers present in v1 train, and reading `unlabeled` yields nothing after a very long wait. Also corrects three things the card got wrong: it referenced three files that were never published (the parquet metadata is blocked by the storage quota), it presented the repacked-audio schema as if it were what you read today, and it did not warn that speaker_id is reused across languages (354 collisions). excluded_rows.csv is added so the two unmapped utterances are documented.c293462 -
remove probe
Verified signaturefafb264 -
probe csv
Verified signatured9c5cac
-
Add ASR v2 speaker-disjoint split definitions
Adds data/ASR_v2/: the complete v2 split assignment for all 19 ASR languages, sharded per language, with the v1-vs-v2 audit tables and full provenance. Purely additive. No existing file is modified and no config is changed, so every current load_dataset call resolves exactly as it did before. Eval-set speaker leakage falls from 97-100% (in 11 of 19 languages) to 0% in every language. The split map reproduces v2 against any copy of v1 without downloading anything further.96ced60 -
Add ASR v2 split summary (probe)
2e3955e